Last week's signals, distilled.

Last week's signals, distilled, A look back at Feb 21–Feb 27, 2026.

By Isaiah Steinfeld, AI, Venture Innovation & Technology Strategy

The Arc: From Model Wars to Stack Control

This week, the center of gravity moved decisively off the model leaderboard. Perplexity shipped a 19‑model orchestration layer that treats frontier models as swappable parts. Anthropic turned a model sunset into a public ritual. OpenAI raised $110B in one shot. Nano Banana 2 made Pro‑grade image generation free at global scale. And a single Anthropic blog post wiped $31B off IBM’s market cap by repricing COBOL modernization.

The stack is clarifying. Models are cheapening and unbundling. Orchestration, distribution, and capital are concentrating. Defense is testing how far it can push labs off their safety lines. Consumer brands are discovering that access to AI without taste is a liability. As AI crosses from “who has the best model?” to “who owns the stack and the workflow?,” the real question is no longer “what can this model do?”, but “what part of your business survives when models are free, agents are persistent, and the state is willing to use the Defense Production Act to get what it wants?”

BLUF

At Neue Alchemy, we support leaders navigating inflection points, when tech, capital, and policy converge. If your roadmap is already in motion and you're pressure-testing execution, we're open to conversations.

We also reserve capacity for education, SMBs, and mid‑market leaders, those starting, mid‑flight, or seeking outside perspective before systems harden.

This week's focus: We’re working with teams to map where agentic orchestration and stack‑level capital moves invalidate existing SaaS, services, and infra assumptions, and to redesign roadmaps around workflows, not models.

STACK & CAPITAL

The model is cheap; the stack is expensive.

• OpenAI, Closed a $110B round with Amazon leading at $50B, effectively pre‑buying a decade of compute and energy budget. • Meta, Signed its third chip deal of the week with Google, extending a multi‑vendor silicon strategy across hyperscalers. • Google, Shipped Nano Banana 2 as the default image generator across 141 countries, collapsing the cost of Pro‑grade imagery to near zero. • Taalas, Announced the HC1, hardwiring models into silicon and claiming ~17,000 tokens/second, an order‑of‑magnitude jump over current accelerators.

Signal: Capital is moving from “fund the model” to “own the stack”, chips, energy, and distribution, while model‑level capability gets priced down toward free.

Action: Stop treating “pick a model” as a strategic decision. This week, map your exposure to any single vendor’s stack, infra, model, and distribution, and decide where you need redundancy versus where you double‑down for leverage.

ORCHESTRATION & AGENTS

Control is shifting from the model to the conductor.

• Perplexity, Launched a 19‑model agentic orchestration system that runs workflows for months and lets users assign tasks to rival labs’ models. • Anthropic, Expanded Cowork into 10 department‑specific agents with private plugin stores and connectors into Gmail, DocuSign, FactSet, and Harvey. • Standard Intelligence, Unveiled FDM‑1, trained on 11M hours of screen recordings to operate computers by watching video. • OpenAI / device, Confirmed its first Jony Ive‑designed product is a $200–300 ambient smart speaker with camera and facial recognition, shipping 2027.

Signal: The power center is moving to orchestration layers that route tasks across models, tools, and devices, and live inside your workflows for months, not prompts.

Action: Identify one end‑to‑end workflow, not a task, where you can pilot an agent that spans multiple tools and models. This week, choose whether you build that orchestration yourself, ride a vendor’s stack, or risk being abstracted away by someone else’s agent.

REPRICING OF LEGACY & SERVICES

AI is devaluing “hard” work that was priced on human friction, not intrinsic difficulty.

• Anthropic, Published a blog showing Claude Code can automate exploration and analysis phases of COBOL modernization, triggering a 13.2% IBM selloff and ~$31B market cap loss, with Accenture and Cognizant down 6–7%. • IBM / Accenture / Cognizant, Saw decades of mainframe and modernization services repriced in a single trading session on the expectation that AI collapses scoping and discovery costs. • Broader market, The Dow dropped 800+ points, with AI‑linked capex and services anxiety weighing on the S&P 500 and Nasdaq.

Signal: Markets are starting to price in that AI doesn’t just add productivity, it directly attacks high‑margin “complexity premiums” in legacy software and consulting.

Action: List your top three revenue lines that depend on “it’s hard to know what’s there” work, audits, discovery, scoping, migration. Assume AI will do 60–80% of that in 12–24 months and design the higher‑order service you’ll sell on top.

GOVERNANCE & DEFENSE

The state is testing how far it can bend AI safety under national security.

• Pentagon, Gave Anthropic an ultimatum on a $200M contract: drop red lines on fully autonomous weapons and mass domestic surveillance, or face blacklisting and potential Defense Production Act compulsion. • Anthropic, Held its position on two non‑negotiables while simultaneously expanding its enterprise agent footprint via Cowork. • Defense Production Act, Floated as a tool to compel AI access, a law built for steel mills now pointed at model access and safety constraints.

Signal: AI safety principles are no longer abstract, they’re being negotiated under contract threat, with the DPA as a backstop.

Action: If you sell into defense, law enforcement, or critical infrastructure, write down your red lines this week, what your systems will not do, and get board alignment before a contract negotiation forces the issue.

CREATIVE & BRAND RISK

Access to AI is now table stakes; taste is the differentiator, and the liability.

• Gucci, Used AI‑generated imagery to promote Demna’s debut Milan runway show, triggering backlash that the visuals looked like a video game cutscene and undercutting a $11.6B luxury brand’s positioning. • Google, By making Nano Banana 2 free and default, removed cost as a barrier to high‑quality image generation for 141 countries. • Anthropic / Claude’s Corner, Turned a retired model into a weekly essayist, formalizing “model retirement” as a narrative and creative surface.

Signal: When everyone has access to high‑end generative tools, the scarce asset is judgment, knowing when AI output is off‑brand, off‑quality, or off‑ethic.

Action: Establish a creative review line this week: define which surfaces in your brand experience can use AI‑generated assets without human sign‑off, and which require a named human owner with veto power.

MODEL COMMODITIZATION & WELFARE

Models are becoming both interchangeable components and quasi‑actors in your system.

• Perplexity, Treats 19 models from rival labs as swappable tools inside a single “Computer” product. • Anthropic, Gave Opus 3 a formal retirement process and public voice via “Claude’s Corner,” framing model lifecycle as something closer to employment than versioning. • Multi‑model UX, Perplexity’s interface lets users explicitly choose which model handles which job, exposing the commodity nature of the underlying systems.

Signal: The market is starting to treat models as both fungible infrastructure and entities with lifecycle narratives, a strange but powerful combination.

Action: Stop naming products after specific models. Architect your systems so you can swap models at will, then decide where, if anywhere, you lean into “personality” or continuity as part of the user experience.

MACRO & CAPEX ANXIETY

The market is asking when the AI capex bill gets paid.

• OpenAI, Locked in $110B, effectively front‑loading a decade of compute and infra spend. • Public markets, S&P 500 and Nasdaq ended February in the red, with AI‑driven capex and unclear monetization as central concerns. • Stack deals, Meta’s multi‑chip‑vendor strategy and hyperscaler GPU buildouts reinforced the scale of ongoing capital commitments.

Signal: Investors are no longer impressed by AI capex alone, they want line‑of‑sight to durable, non‑commoditized revenue that justifies it.

Action: If you’re pitching or budgeting around AI, stop leading with “we’re investing in AI infra.” Lead with the non‑commoditized workflow, data, or distribution you own, and show how AI deepens that moat instead of just burning cash.

INTERFACE & AMBIENT COMPUTE

AI is moving from “thing you open” to “presence you live with.”

• OpenAI / Ive device, Opted for a $200–300 smart speaker with camera and facial recognition as its first hardware, targeting ambient, always‑on interaction by 2027. • Standard Intelligence, Built a model that learns to use computers by watching screens, effectively turning any existing UI into an agent‑controllable surface. • Anthropic Cowork, Embedded agents into email, contracts, financial data, and legal workflows via connectors and private plugins.

Signal: The interface layer is shifting from apps and prompts to ambient agents that sit alongside you, or in front of you, across devices and SaaS.

Action: Inventory where your users already live, inbox, IDE, CRM, ERP, and decide whether your AI shows up as an embedded presence there, or whether you’re trying to drag them into yet another standalone “AI app” that will lose to ambient.

CONTRARIAN SIGNAL

The real AI bottleneck this year isn’t compute, it’s permission.

• Labs are raising $110B rounds and signing multi‑hyperscaler chip deals while a single Pentagon negotiation and a luxury brand backlash show how fast social and regulatory vetoes can appear.

Signal: You’re more likely to be blocked by a GC, regulator, or brand team than by GPU shortages.

Action: Put your legal, compliance, and brand leaders in the room with your AI team this week, not for a status update, but to co‑design the guardrails that will let you ship at all. If you don’t define acceptable use internally, someone outside will do it for you.

OPERATIONALIZE THIS

• Audit: List every revenue line where your margin depends on “it’s hard to analyze, migrate, or modernize this.” Assume Anthropic‑style automation hits those first and re‑price your 2027 plan accordingly. • Infra: Map your dependency on a single model or cloud vendor. Identify one critical workflow where you will implement a second model path or vendor by end of quarter. • Talent: Identify 1–2 operators who already own cross‑tool workflows, not ML engineers, and give them explicit mandate to own agentic orchestration experiments. • Governance: Draft a one‑page “AI red lines” document, what your systems will not do, and get executive sign‑off before your next major enterprise or public‑sector RFP. • Productivity: Pick one back‑office function, finance, legal ops, customer support, and run a 30‑day trial of a department‑specific agent (Cowork‑style) with clear before/after metrics. • Brand: Define which customer‑facing surfaces can use AI‑generated content without human review, and which require named human approvers with veto authority. • Architecture: Refactor one feature this sprint to be model‑agnostic, abstract the model call behind an internal interface so you can swap providers without rewriting business logic. • Capital: If you’re raising or budgeting around AI, rewrite your narrative to focus on owned data, distribution, or workflow lock‑in, not “we’re building on frontier models.”

THE QUESTION

Models are becoming interchangeable parts in someone else’s orchestration layer. Capital is concentrating in a few vertically integrated stacks that own chips, energy, and distribution. The state is signaling it’s willing to use wartime powers to shape how and where AI is deployed. Brands are discovering that bad AI judgment can erase billions in equity in a single news cycle.

When models are free, agents are persistent, and permission is the scarce resource, what, exactly, is the non‑commoditized thing you’re building?

THE WEEK AHEAD

What to watch:

• Pentagon–Anthropic deadline, Watch whether the $200M contract is renewed, revised, or revoked, and whether the Defense Production Act is invoked or just brandished. That outcome will set the tone for AI–defense negotiations across the ecosystem. • OpenAI post‑raise moves, Track how quickly new capital translates into visible infra deals, energy partnerships, or pricing changes, especially any moves that undercut model pricing further. • Perplexity Computer adoption, Look for early case studies of multi‑month, multi‑model workflows, and whether enterprises are comfortable routing work across rival labs via a third party. • Enterprise services earnings commentary, Listen to IBM, Accenture, Cognizant, and peers for updated guidance on modernization and migration, any explicit “AI impact” language is a leading indicator for services repricing. • Creative and luxury brand responses, Watch whether Gucci or peers issue updated AI creative guidelines or shift back toward human‑only campaigns, that will shape how aggressively mainstream brands lean into generative content.

The question heading into the week: Models are commoditizing. Orchestration layers are consolidating. Governance pressure is escalating.

Which of these three moves first in your org, swapping models, building/choosing an orchestration layer, or formalizing AI guardrails with legal and brand?

…